generated: '2026-08-24'
method: searched
source: >-
  https://www.agentsea.dev/docs/mcp-overview/, https://www.agentsea.dev/docs/mcp-servers/,
  https://www.agentsea.dev/docs/mcp-tools/, https://www.agentsea.dev/docs/acp-integration/,
  https://www.agentsea.dev/docs/gateway/, https://www.agentsea.dev/api/,
  https://hex.pm/packages/agentsea_mcp
note: >-
  AgentSea's conformance posture is that of a CLIENT and an ADAPTER, not a server. It speaks MCP to
  other people's servers, speaks the Agentic Commerce Protocol to merchants' endpoints, and
  re-implements OpenAI's Chat Completions contract so that OpenAI SDKs can talk to it. Recording
  the direction matters: an MCP client is not an MCP server, and only the latter is an agent
  surface a third party can call.
standards:
- id: model-context-protocol
  name: Model Context Protocol
  conforms: true
  role: client
  evidence: >-
    First-class MCP client with a server registry (mcpRegistry.addServer) supporting stdio
    transport, documented against the official Anthropic servers (filesystem, github, postgres,
    slack, puppeteer, gdrive). The Elixir library agentsea_mcp is described on Hex as "a Model
    Context Protocol client with stdio and streamable-HTTP transports". AgentSea publishes no MCP
    server of its own - see mcp/agentsea-mcp.yml.
  docs: https://www.agentsea.dev/docs/mcp-overview/
- id: agentic-commerce-protocol
  name: Agentic Commerce Protocol (ACP)
  conforms: true
  role: client
  evidence: >-
    ACPClient plus createACPTools() in @lov3kaizen/agentsea-core give an agent 14 commerce tools -
    product search/browse/details/recommendations, cart add/update/remove/view, checkout
    initiate/validate/complete, order track/history/details - against a merchant-supplied ACP
    baseUrl with an apiKey and merchantId. Payment is delegated to Stripe/PayPal. AgentSea is the
    buyer-side client; it does not serve an ACP endpoint.
  docs: https://www.agentsea.dev/docs/acp-integration/
- id: openai-chat-completions
  name: OpenAI Chat Completions API compatibility
  conforms: true
  role: server (self-hosted)
  evidence: >-
    @lov3kaizen/agentsea-gateway serves POST /v1/chat/completions and is documented as usable with
    any OpenAI SDK; the Elixir agentsea_web Phoenix router declares the same route
    (scope "/v1" ... post "/chat/completions", ChatController, :create). Virtual model aliases
    "best", "cheapest" and "fastest" are AgentSea extensions on top of the OpenAI request shape.
  docs: https://www.agentsea.dev/docs/gateway/
- id: sse
  name: Server-Sent Events (HTML Living Standard / text-event-stream)
  conforms: true
  role: server (self-hosted)
  evidence: >-
    POST /agents/:name/stream streams with Accept text/event-stream; documented event types are
    iteration, content, tool_calls, tool_result, done and error. The Elixir provider implements
    real SSE streaming against Anthropic.
  docs: https://www.agentsea.dev/api/
- id: json-schema
  name: JSON Schema (via Zod)
  conforms: true
  role: internal
  evidence: >-
    Tool definitions carry an inputSchema expressed as a Zod schema; @lov3kaizen/agentsea-structured
    enforces JSON Schema / Zod validation on structured outputs with automatic retries, and
    SchemaGuard validates output against a Zod schema.
  docs: https://www.agentsea.dev/docs/structured-output/
- id: oauth2
  name: OAuth 2.0
  conforms: false
  evidence: >-
    No OAuth anywhere. /.well-known/oauth-authorization-server and
    /.well-known/oauth-protected-resource both return 404 on agentsea.dev and www.agentsea.dev.
    Tenant auth is a scoped, SHA256-hashed API key minted by the self-hoster.
- id: rfc9457
  name: RFC 9457 Problem Details for HTTP APIs
  conforms: false
  evidence: No problem+json envelope or error catalogue is published for either HTTP surface.
- id: a2a
  name: A2A (Agent2Agent) Agent Card
  conforms: false
  evidence: >-
    /.well-known/agent-card.json and /.well-known/agent.json probed on both agentsea.dev and
    www.agentsea.dev on 2026-08-24 - all four returned HTTP 404.
domain_standard:
  market: agent development kits / agentic AI frameworks
  declared: partial
  detail: >-
    The agent-framework market has no single ratified interoperability standard, and the checks that
    exist in the rubric read a served contract rather than a prose claim. AgentSea's strongest
    contract-level signature is the OpenAI Chat Completions shape on /v1/chat/completions, which is
    visible in open source in the Elixir router and in the gateway docs. MCP and ACP are real and
    substantial, but AgentSea sits on the client side of both, so they are recorded as role: client
    rather than as a served surface. REWARD-ONLY: nothing is asserted here that the published code
    or docs do not state.
compliance:
  certifications: []
  programs: []
  note: >-
    No SOC 2, ISO 27001, PCI, HIPAA or FedRAMP claim; no trust center; no security.txt; no
    SECURITY.md in the repository (probed 2026-08-24, HTTP 404). No Compliance pointer is emitted
    in apis.yml. The nearest published security artefacts are in-product, not organizational: the
    guardrails package (PII detection, prompt-injection, jailbreak and data-leakage guards), the
    redteam package with hash-chained audit storage, and the v0.8.0/v1.2.0 release notes recording
    CVE remediation and a pnpm audit going from 168 known vulnerabilities to zero.